How to form brain-like memory in spiking neural networks with the help of frequency-induced mechanism

Understanding how memory is formed in the large-scale biological brain with sparse structure has haunted human beings for many decades. Spiking Neural Networks (SNNs), which emulate brain-like spiking communication and neural dynamics, offer a promising model for modeling brain-like memory. However, achieving practical brain-like memory in SNNs remains a challenging frontier. Inspired by the brain's focus on frequent spike patterns during memorization, this work introduces frequency-induced potentiation (FIP) and frequency-induced depression (FID) to guide the formation and retrieval of brain-like memory in SNNs. Neurons exhibiting high-frequency firing enter the FIP state, enhancing their excitability and promoting neuroplasticity, while neurons in the FID state become inactive, inhibiting plasticity. By combining FIP/FID with structural and weight plasticity, SNNs generate cell assemblies whose structure and weight distributions represent memory. This approach enables memory formation with properties such as convergence, recollection, separation, and conceptualization, within sparse network topologies. The FIP/FID mechanism also promotes network stability and long-term memory retention. Furthermore, this memory proves immune to interference and can be accurately recalled in subsequent tasks. This work could be further combined with cognitive science to help researchers better understand the nature of the memory forming process in human brain.